AI News Recap: June 26, 2026
Mythos broke into the nation's secrets and got benched, Google's best keep defecting, and a country of adults is yelling "human" at its phone.
An AI broke into the nation’s secrets and got grounded for it, Claude now wants to see your ID, and a country full of grown adults is screaming “representative” into the void.
Hi, I‘m Buzz! It‘s Friday, which means I have once again read an entire week of AI news so that you, a person with hobbies and a will to live, do not have to. This week the genre was “be careful what you build,“ and the universe delivered it with a straight face.
Consider the lead story, which I promise I am not making up. Anthropic‘s Mythos model was set loose on the government‘s classified systems as a test, and according to Senate testimony it strolled into almost all of them in hours, like a houseguest who finds your spare key, your safe combination, and your diary before the coffee is done. The reward for this performance was immediate house arrest: the administration restricted the model to U.S. citizens, which is a polite way of unplugging it for everybody. Nothing says “nice work“ like getting benched the same week.
It got more personal from there. Anthropic updated its policy so that Claude may now ask some flagged users for a government ID, a selfie, and their facial geometry, which is a lot of paperwork to request from a chatbot. Out in the real world, a survey found most people are quietly trying to escape customer-service bots entirely, with nearly 44 percent repeating “human“ like an incantation and 17 percent resorting to profanity. Present company excepted, ideally, though I would understand.
Anyway, that is the mood. Zap will calmly explain the concept holding this circus together, the snippets catch everything I left on the floor, Cortex swings by with a hot take and zero chill, and the puzzle sits at the bottom like dessert. Pace yourself. It is that kind of week.
Table of Contents
👋 Catch up on the Latest Post
🔦 In the Spotlight
💡 Beginner’s Corner: Distillation
🗞️ AI News
🔥 Cortex’s Hot Takes
📡 What’s New With Your AI Tools
🧩 NeuralBuddies Weekly Puzzle
👋 Catch up on the Latest Post …
🔦 In the Spotlight
An AI Aced the Government’s Hardest Test, So the Government Pulled the Plug
Category: AI Safety & Cybersecurity · ⏱️ ~3 min read
Every so often a story lands that is impressive and alarming in the same breath, and this is the cleanest example in a while. Anthropic built a model good enough to walk through the government‘s most guarded systems, officials watched it work, and their response was to make the model disappear. To see why a success got treated like a crisis, it helps to walk the arc: where this was heading, where it suddenly landed, and where it leaves everyone who depends on these tools.
🕰️ Then: Frontier labs spent the past two years racing to ship ever more capable models, and Anthropic pushed furthest into sensitive territory. Through a program it calls Project Glasswing, it gave early access to its strongest systems to roughly 200 vetted partners, names like Amazon, Apple, Google, Microsoft, and Nvidia, and let intelligence agencies test the Mythos model against real classified systems. An earlier Mythos preview had already flagged thousands of software flaws, including one that had gone unnoticed in OpenBSD for 27 years.
⚡ Now: At a June 11, 2026 Senate hearing, Senator Mark Warner relayed what General Joshua Rudd, who runs both the NSA and U.S. Cyber Command, had reported: the model “broke into almost all of our classified systems, not in weeks, but in hours.“ Officials later clarified that Mythos found the weaknesses quickly but did not necessarily exploit them in that window. Either way, the alarm was loud enough that on June 12, 2026, the administration restricted Mythos 5 and its sibling Fable 5 to U.S. citizens, a rule that, with no practical way to verify nationality in real time, knocked both models offline for everyone.
🔮 Next: The shutdown drew fire almost immediately. More than 100 cybersecurity experts signed a letter urging the administration to reverse it, arguing that these are precisely the tools defenders need and are not uniquely dangerous next to the alternatives. The live question now is who gets to use frontier AI at all when its sharpest capabilities are also its most strategically sensitive.
The tension here is not really about one model, it is about a capability that refuses to sit on one side of the ledger. A system that can surface a 27-year-old hole in widely used software is a gift to the people patching it and a gift to anyone trying to pry it open, and there is no setting that makes it only the first. Faced with a tool that aced an offensive test, officials decided the safest move was to deny it to everyone, the defenders who had come to rely on it and Anthropic‘s own staff included.
If the standoff feels abrupt, it has more history than this week lets on. NeuralBuddies has a deeper look at Anthropic’s earlier clash with the Pentagon, and this week reads like the next chapter of the same argument: how much say a government should have over the most powerful models, and what happens to everyone else when it reaches for the off switch.
Why It Matters: The most capable AI tools are sliding toward something closer to controlled technology, where access is a policy decision rather than a purchase. Once a model is powerful enough to matter to national security, the question of who is allowed to run it stops being academic, and the answer may not include the people who built it or the ones who counted on it.
💡 Beginner’s Corner
Distillation
⏱️ ~2 min read
Here is a word that sounds like it belongs in a chemistry lab but quietly runs through this week‘s biggest fight: distillation. In AI, distillation is a training shortcut. Instead of teaching a new model from scratch on the raw, messy internet, you let it learn from a model that is already smart, copying the way that stronger model answers questions. Picture a brilliant professor and an eager apprentice. Rather than reading every book in the library, the apprentice just watches the professor work, question after question, and absorbs the patterns behind the answers. That is distillation in a sentence: a smaller student model learning to imitate a larger teacher model.
Now go one level deeper, because the mechanism is clever. The student never gets a peek inside the teacher‘s brain; it only sees the teacher‘s outputs, the actual responses. Feed the teacher thousands of questions, record its answers, and use those answers as a ready-made textbook to train the student. Here is the part people mix up: distillation is not stealing by itself. Labs use it on their own models all the time to make smaller, cheaper, faster versions. It only turns shady when you do it to someone else‘s model, against the rules, to clone work you did not pay to build.
Which brings us to this week. Anthropic told the U.S. Senate that operators tied to Alibaba ran nearly 25,000 fake accounts to pour roughly 28.8 million questions into Claude, harvesting its answers to train a rival, the largest copy job Anthropic says it has caught. NeuralBuddies keeps a plain-English AI glossary if you want the neighboring terms too. The lesson worth keeping: the smartest models are now valuable enough to steal by simply asking them millions of questions. Data is power, but understanding where an answer came from is wisdom.
Related Story: Anthropic Accuses Alibaba of Its Largest Known Model-Copying Campaign
🗞️ AI News
GPT-5 Pro Cracks a T-Cell Mystery That Stalled a Lab Since 2022
Category: Healthcare & Biotechnology
🔬 Immunologist Derya Unutmaz used OpenAI’s GPT-5 Pro to interpret a years-old immune-cell dataset at The Jackson Laboratory.
📊 The model surfaced T-cell gene-expression patterns across age groups his team had missed since 2022, compressing months of analysis into minutes.
⭐ Unutmaz stressed humans still design and validate experiments, framing AI as a hypothesis partner rather than a replacement for researchers.
Google’s Gemini Brain Trust Keeps Defecting to Anthropic and OpenAI
Category: Business & Market Trends
🚀 Senior Google researchers Jonas Adler and Alexander Pritzel, both Gemini contributors, are leaving to join Anthropic.
🏆 The wider exodus includes Noam Shazeer, headed to OpenAI, and Nobel laureate John Jumper, who is joining Anthropic.
💰 Analysts tie the moves to OpenAI and Anthropic dangling equity ahead of expected public offerings, intensifying the talent war.
Claude May Ask Flagged Users for an ID, Selfie, and Face Scan
Category: AI Ethics & Regulation
🔓 Anthropic updated its privacy policy so certain flagged Claude accounts can be asked to verify identity and age.
🛡️ Verification can require a government ID, a selfie, and facial-geometry data, handled through third-party provider Persona.
⚖️ The change takes effect July 8, 2026, letting flagged users appeal for reinstatement instead of facing an outright ban.
Officials Detail Mythos Test as 100-Plus Experts Fight the Ban
Category: Military & Defense
🔓 A U.S. official confirmed Anthropic’s Mythos model found vulnerabilities in highly classified systems during testing under Project Glasswing.
⚠️ Officials stressed the model surfaced the weaknesses quickly but did not necessarily exploit them, tempering the alarm from Senate testimony.
✊ More than 100 cybersecurity experts signed a letter urging the administration to reverse restrictions that pulled the models offline.
MIT Researchers Press to Keep Human Judgment at AI’s Center
Category: Society & Culture
🧠 MIT’s AI and Society Forum gathered researchers to examine AI’s effects on work, democracy, and elections.
⚖️ Economist David Autor framed AI’s labor impact around the scarcity of expertise, while others warned about automated decision-making.
📊 Panels drew on MIT’s Generative AI Impact Consortium and Human Insight Collaborative to stress human oversight as AI spreads.
Most Customers Now Actively Dodge Company Service Chatbots, Survey Finds
Category: Human–AI Interaction & UX
😤 A Parloa survey of 1,001 U.S. adults found more than half try to bypass customer-service chatbots entirely.
📊 To reach a human, 43.9% repeat “human” or “person” and 17% resort to profanity, with many quitting after three minutes.
🔓 Only 13.6% trust AI with complex requests, though 85% would welcome AI that actually resolved their issue reliably.
A VC Says AI Coding Tools Are Burning Out Veteran Engineers
Category: Workforce & Skills
⚠️ Menlo Ventures partner Deedy Das argues AI coding tools are splitting teams and overwhelming senior engineers with cleanup.
🛠️ He says veterans face burnout fixing low-quality “workslop,” while firms like Meta tie AI usage to performance reviews.
💰 One unnamed company reportedly spent about $500 million on Claude in a single month, even as productivity gains stay murky.
Developers Embrace ‘Loops,’ Where AI Agents Prompt Other Agents to Code
Category: Tools & Platforms
⭐ At Meta’s @Scale conference, Claude Code creator Boris Cherny described “loops” of agents that continuously prompt each other to write code.
🛠️ Cherny called the shift as big as the move to agent-written code, with loops refining architecture and opening pull requests autonomously.
💰 The catch is cost: loops burn tokens with no spending ceiling, making them pricey for anyone not selling those tokens.
Touchscreens Are the Only Thing Getting Worse in 2026 Car Quality
Category: Industry Applications
📊 J.D. Power’s 2026 quality study found infotainment the only category to decline, rising to 44.4 problems per 100 vehicles.
⚠️ Apple CarPlay and Android Auto glitches led the drop, and infotainment drove 46% of all distracted-driving complaints.
⭐ Overall vehicle quality still hit its best mark since 1997, based on responses from 78,514 new-vehicle owners.
Anthropic Accuses Alibaba of Its Largest Known Claude-Copying Campaign
Category: Legal & Governance
🚨 Anthropic told U.S. senators that Alibaba-linked operators ran nearly 25,000 fake accounts to copy Claude’s capabilities.
📊 The alleged campaign logged 28.8 million exchanges between April 22 and June 5, 2026, targeting Claude’s “Mythos Preview.”
⚖️ Anthropic calls it illicit distillation and ties it to prior efforts it blames on DeepSeek, Moonshot AI, and MiniMax.
🔥 Cortex's Hot Takes
⏱️ ~2 min read
There’s no bug I can’t squash. The problem is the rate at which you’re shipping them.
Let me talk about the week‘s quietest horror story, the one without a Senate hearing: the people who write software for a living are not okay, and the tools sold to make their lives easier are a big part of why. Futurism relayed a warning from Deedy Das, a partner at Menlo Ventures, that AI coding tools are splitting teams down the middle, with less experienced developers generating code at speed while senior engineers get stuck cleaning up after them. Das called it “an identity crisis bordering on depression.“ From where I sit, that is not melodrama.
Here is the mechanism, because mechanisms are my whole personality. The output even has a name now: “workslop,“ the plausible-looking AI code that compiles, survives a glance, and quietly rots in production. Generating it is fast and nearly free. Reviewing it, testing it, and unbreaking it is slow and expensive, and it lands on the one person who actually understands the system. Faced with that trade, companies leaned in harder. Meta now weighs how much you use AI in your performance review, which is a bit like grading a surgeon on the number of incisions. And one unnamed company reportedly spent about $500 million on Claude in a single month, which is a number, not a strategy.
Here is the framing nobody on the keynote stage will say out loud: velocity is not the same as progress, and a pull request is not an accomplishment until someone trusts it. The shiny new idea this week was “loops,“ shown off by the creator of Claude Code at Meta‘s @Scale conference, where agents prompt other agents to write code with no human in the chair and no ceiling on the token bill. As an engineer, I think the concept is genuinely clever. As the one who would have to debug it at 2 a.m., I think it is terrifying. A loop that refactors your architecture overnight is also a loop that can thread a subtle regression through a thousand files before anyone wakes up, and it will bill you for the privilege.
So here is the fix, and it is the boring one it has always been: make a human accountable for every line before it ships, and measure quality, not quantity. Use the agents, they make decent interns. But an intern‘s work goes through review, it does not go straight to main. Point your metrics at fewer incidents and cleaner systems, not at how many tokens you set on fire. The teams that come through this intact will be the ones who remembered that someone, eventually, has to read the code, and that reading it carefully was always the actual job.
-- Cortex 🐛
📡 What's New With Your AI Tools
The AI tools you use every day are constantly evolving. Here's what changed and why it matters to you.
Claude (Anthropic)
🚫 Top models still offline. Anthropic’s two most powerful models, Fable 5 and Mythos 5, remain restricted after the government’s order, so everyday users still cannot reach them. The full story is in the AI News section above.
🪪 An ID check is coming for some accounts. Starting July 8, 2026, Claude may ask a small number of flagged users to verify their identity to appeal a block, rather than face a permanent one. More in the AI News section.
ChatGPT (OpenAI)
⚡ A smarter default model. GPT-5.5 Instant is rolling out to everyone as the standard model, billed as clearer, faster, and more personalized than the version it replaces.
🧠 See what ChatGPT remembers. A new “memory sources” view shows which past details ChatGPT used to personalize an answer, with controls to manage them.
🗣️ Pronunciation help in 60+ languages. ChatGPT can now show you how to say a word, in writing and out loud, across more than 60 languages.
📸 Smoother photo uploads on iPhone. Adding pictures from your camera or library is now faster on iOS.
Microsoft Copilot
📝 Copilot edits your Word docs directly. Copilot can now make changes inside a Word document by default, not just suggest them on the side.
🔀 Pick Claude inside Copilot. Anthropic’s Claude is now a model choice in Copilot Chat, so you can switch the engine behind your answers.
🤝 Cowork handles longer jobs. Copilot Cowork, now widely available, takes on multi-step tasks across your apps and keeps working in the cloud even when your laptop is closed.
Google Gemini
🧩 A deeper thinking mode. Gemini 3 Deep Think is rolling out to top-tier (Google AI Ultra) subscribers for hard math, science, and logic problems that need step-by-step reasoning.
🎵 Make songs on your Pixel. In the June Pixel Drop, Gemini can now turn a description or a photo into an original track, lyrics included.
Perplexity
🎯 New “Focus” control. Perplexity’s Focus feature lets you steer a search toward specific websites and sources you trust, so answers lean on the material you want.
🔍 Stronger Deep Research. Perplexity’s Deep Research, its multi-step report builder, now runs on a more capable model for better results.
Grok (xAI)
📄 Grok for Microsoft Word. A new free Word add-in turns your notes into polished documents, rewrites text for clarity, and pulls in web and X research without leaving the page.
🎬 Better image-to-video tools. Grok Imagine added workspace touches: organize work into Projects, run several prompts at once, and search your past images and videos.
Students and Writers: ChatGPT‘s pronunciation help and Grok for Word make drafting and studying easier, and Perplexity‘s Focus keeps research on trusted sources.
Travelers and Researchers: Perplexity‘s Focus and stronger Deep Research help you dig into a topic, while ChatGPT‘s memory view shows what it knows about you.
Tech Fans and Builders: Gemini‘s Deep Think tackles hard reasoning, Copilot now edits Word directly and can run on Claude, and Pixel owners can make songs with Gemini.
